Buckle up: Nvidia is «reinventing the personal computer,» CEO Jensen Huang announced during the keynote at the company’s GTC session at Computex in Taipei on Monday. Microsoft and Nvidia have been tightening their partnership ahead of the eagerly awaited launch of the RTX Spark.
The RTX Spark is a new Arm‑based system‑on‑chip (SoC) platform that brings Nvidia’s Blackwell architecture to thin, light Windows laptops and mini‑desktops. Its aim is to deliver desktop‑grade compute power for personal agents, creative workloads and gaming, while avoiding the bulk, power draw and cooling demands of traditional discrete GPUs.
This chip joins Qualcomm’s Snapdragon X line that also runs Windows on Arm, with similar promises of «all‑day battery life.» While Snapdragon chips achieve that, the Nvidia solution is built for far more demanding tasks than the Snapdragon family.
We’re not talking about casual browsing – the Spark is meant to handle massive 90 GB‑plus 3D scenes, 12K 4:2:2 video editing, 4K AI video generation, 120‑billion‑parameter LLMs with up to a million‑token context running locally, and AAA gaming at 1440p and well over 100 fps, all of which can quickly drain a battery. Whether the Spark can sustain such workloads in everyday use remains to be seen.
This is the first of what Nvidia describes as a family of chips spanning several price tiers. The initial devices slated for a fall release include:
- Microsoft Surface Laptop Ultra
- Dell XPS 16
- Asus ProArt P14 and P16
- HP Omnibook X 14, Omnibook Ultra 16
- Lenovo Yoga Pro 9n
- MSI Prestige N16 Flip AI
The 15‑inch Surface Laptop Ultra stands out because Microsoft hasn’t refreshed its displays in ages, and previous Surface models never featured discrete GPUs despite their price tags. The Ultra sports a high‑resolution (262 ppi) 15‑inch mini‑LED touchscreen with HDR support and a peak brightness of 2,000 nits – a big step up from the older, lackluster version. Microsoft’s Surface Laptop Studio has been unchanged for three years, and the RTX Spark plus the new screen could be the upgrade it needs to revive the line.
Mini‑desktops are also in the pipeline, a segment that appears to be enjoying a resurgence thanks to developer interest. RTX Spark‑powered mini‑desktops will compete with AMD’s Ryzen AI Halo‑based offerings and are expected from manufacturers such as Acer, Asus, Dell, HP and Lenovo.
Nvidia plans to deliver a desktop, laptop and workstation for each chip generation. Pricing is still uncertain due to the current volatility in component costs – AI’s voracious appetite for chips, memory and SSDs has created shortages that are driving up prices and limiting configuration options.
Spark it up
The chip derives from the DGX Spark (GB10), which powers Linux‑based compact desktops aimed at developers and now the Windows‑based DGX Station. Co‑designed with MediaTek, it mirrors the DGX’s specs: 6,144 CUDA cores, a 20‑core Grace CPU, up to 128 GB of RAM, and support for agents up to 120 B parameters with a 1 M token context. (For comparison, AMD says its top Ryzen AI Max Pro 400 series can handle models of up to 300 B parameters.)
GPU‑wise it lines up roughly with an RTX 5070, but the unified memory architecture gives it far more than the 12 GB VRAM of a typical 5070. Nvidia says configurations could start at just 16 GB of system RAM, which might become a bottleneck in scenarios where a dedicated 5070 with 12 GB VRAM would not.
The reference gaming performance is quoted at 100 fps at 1440p, though it’s unclear whether that figure assumes DLSS 4.5 is enabled. Nvidia claims an overall AI throughput of one PFLOPS, based on FP4 calculations – a data format that offers a good balance of speed and accuracy, and the first consumer SoC to support it in hardware.
The main rivals are Apple’s M5 Pro and M5 Max MacBook Pros, which target the same professional audience but lack FP4 and FP8 support, potentially limiting their AI capabilities.
Power draw can vary widely, from “single‑digit” watts up to 80 W, so actual performance – especially on battery – will depend heavily on how OEMs tune the chip. For reference, Intel’s Core X9‑388H spans a 15 W‑85 W envelope.
The Spark also includes an NPU, which Nvidia keeps quiet about, but devices featuring the chip qualify for Copilot Plus, meaning they must achieve at least 40 TOPS.
While the RTX Spark looks formidable, Nvidia continues to keep a clear line between its professional and consumer divisions, opting not to run a certification program for applications or support ECC memory.
Windows updates for Arm
Microsoft is working alongside Nvidia to roll out the Windows changes needed to fully exploit the new SoC. Like Qualcomm’s Snapdragon X series, Windows does not natively run Arm instructions the way it does x86; instead it relies on an emulation layer called Prism to translate code.
Supporting the RTX Spark required updating Prism and other core Windows components to efficiently spread work across CPU cores, balance thermals, and manage the larger unified memory pool for GPU‑accelerated AI (TensorRT). Nvidia’s deeper investment in Windows gaming also means collaboration on anti‑cheat compatibility (e.g., Epic’s Easy Anti‑Cheat) and Xbox app integration.
Adobe is re‑engineering parts of its imaging stack to tap the Spark directly, adding new pipelines that accelerate GPU‑ and AI‑heavy features such as complex timeline rendering in Premiere Pro and smarter brushes in Photoshop. While CUDA and TensorRT already run on Nvidia’s discrete mobile GPUs, extracting full performance from the Spark’s architecture requires additional work.
Nvidia is also bringing its OpenShell security framework to Windows, allowing users to set guardrails for agents, route queries to approved local models based on privacy policies, and mask personal data when calling cloud models. These controls will be showcased at Microsoft’s Build conference in early June.
By extending everyday agent capabilities beyond developers, Nvidia hopes to overcome the current trust barrier that limits widespread adoption of on‑device agents. OpenShell will be integrated into the existing agent platforms OpenClaw and Hermes.
